The ASSET intercomparison of ozone analyses: method and first results

Transcription

1 Atmos. Chem. Phys., 6, , 26 Author(s) 26. This work is licensed under a Creative Commons License. Atmospheric Chemistry and Physics The ASSET intercomparison of ozone analyses: method and first results A. J. Geer 1,1, W. A. Lahoz 1, S. Bekki 2, N. Bormann 3, Q. Errera 4, H. J. Eskes 5, D. Fonteyn 4, D. R. Jackson 6, M. N. Juckes 7, S. Massart 8, V.-H. Peuch 9, S. Rharmili 2, and A. Segers 5 1 Data Assimilation Research Centre, University of Reading, Reading, UK 2 CNRS Service Aeronomie, Université Pierre et Marie Curie, Paris, France 3 European Centre for Medium-Range Weather Forecasts, Reading, UK 4 Institut d Aéronomie Spatiale de Belgique, Brussels, Belgium 5 Royal Netherlands Meteorological Institute, De Bilt, The Netherlands 6 Met Office, Exeter, UK 7 British Atmospheric Data Centre, Rutherford Appleton Laboratory, Chilton, nr Didcot, UK 8 CERFACS, Toulouse, France 9 CNRM-GAME, Météo-France and CNRS URA 1357, Toulouse, France 1 Now at European Centre for Medium-Range Weather Forecasts, Reading, UK Received: 6 February 26 Published in Atmos. Chem. Phys. Discuss.: 7 June 26 Revised: 4 October 26 Accepted: 28 November 26 Published: 5 December 26 Abstract. This paper aims to summarise the current performance of ozone data assimilation (DA) systems, to show where they can be improved, and to quantify their errors. It examines 11 sets of ozone analyses from 7 different DA systems. Two are numerical weather prediction (NWP) systems based on general circulation models (GCMs); the other five use chemistry transport models (CTMs). The systems examined contain either linearised or detailed ozone chemistry, or no chemistry at all. In most analyses, MIPAS (Michelson Interferometer for Passive Atmospheric Sounding) ozone data are assimilated; two assimilate SCIAMACHY (Scanning Imaging Absorption Spectrometer for Atmospheric Chartography) observations instead. Analyses are compared to independent ozone observations covering the troposphere, stratosphere and lower mesosphere during the period July to November 23. Biases and standard deviations are largest, and show the largest divergence between systems, in the troposphere, in the upper-troposphere/lower-stratosphere, in the upperstratosphere and mesosphere, and the Antarctic ozone hole region. However, in any particular area, apart from the troposphere, at least one system can be found that agrees well with independent data. In general, none of the differences can be linked to the assimilation technique (Kalman filter, three or four dimensional variational methods, direct inversion) or Correspondence to: A. J. Geer (alan.geer@ecmwf.int) the system (CTM or NWP system). Where results diverge, a main explanation is the way ozone is modelled. It is important to correctly model transport at the tropical tropopause, to avoid positive biases and excessive structure in the ozone field. In the southern hemisphere ozone hole, only the analyses which correctly model heterogeneous ozone depletion are able to reproduce the near-complete ozone destruction over the pole. In the upper-stratosphere and mesosphere (above 5 hpa), some ozone photochemistry schemes caused large but easily remedied biases. The diurnal cycle of ozone in the mesosphere is not captured, except by the one system that includes a detailed treatment of mesospheric chemistry. These results indicate that when good observations are available for assimilation, the first priority for improving ozone DA systems is to improve the models. The analyses benefit strongly from the good quality of the MIPAS ozone observations. Using the analyses as a transfer standard, it is seen that MIPAS is 5 higher than HALOE (Halogen Occultation Experiment) in the mid and upper stratosphere and mesosphere (above 3 hpa), and of order 1 higher than ozonesonde and HALOE in the lower stratosphere (1 hpa to 3 hpa). Analyses based on SCIA- MACHY total column are almost as good as the MIPAS analyses; analyses based on SCIAMACHY limb profiles are worse in some areas, due to problems in the SCIAMACHY retrievals. Published by Copernicus GmbH on behalf of the European Geosciences Union.

2 5446 A. J. Geer et al.: Intercomparison of ozone analyses 1 Introduction The Assimilation of ENVISAT Data project (ASSET, http: //darc.nerc.ac.uk/asset) aims to provide analyses of atmospheric chemical constituents, based on the assimilation of observations from ENVISAT, and to develop chemical weather and UV forecasting capabilities. Data are assimilated into a variety of different systems, including chemical transport models (CTMs) with detailed chemistry or linearised chemistry, and Numerical Weather Prediction (NWP) systems based on General Circulation Models (GCMs), either with linearised chemistry, or coupled to detailedchemistry CTMs. Data assimilation techniques (see e.g., Kalnay, 23) include three and four-dimensional variational data assimilation (3-D-Var and 4-D-Var) and the Kalman Filter (KF). It is hoped that, by confronting these various models and techniques with the newly available ENVISAT observations, it will be possible both to gain an understanding of their strengths and weaknesses, and to make new developments. A number of ozone analyses have been created within the AS- SET project; this paper compares them to independent observations and to ozone analyses from outside the project. The aim is to summarise the current performance of ozone data assimilation systems, to show where they can be improved, and to quantify their errors. Datasets of assimilated ozone will be useful for research and monitoring of ozone depletion (e.g., WMO, 23), tropospheric pollution, and UV fluxes, and beyond this, ozone assimilation is expected to bring a number of benefits in NWP. First, in the upper-troposphere/lower-stratosphere (UTLS), ozone has a photochemical relaxation time of order 1 days, and it can be used as a tracer to infer atmospheric motions using 4D-Var (e.g., Riishøjgaard, 1996; Peuch et al., 2). Second, NWP systems have typically used a zonal mean ozone climatology in modelling heating rates and in the forward radiative transfer calculations used in the assimilation of satellite radiances such as those from the Atmospheric Infrared Sounder (AIRS). An estimate of the true 3-D ozone distribution is likely to improve these calculations. Experiments at ECMWF found that variations in ozone amounts of 1 to 2 could result in changes in modelled tropical UTLS temperatures of up to 4 K (Cariolle and Morcrette, 26). The diurnal cycle of ozone is important in the middle atmosphere. Model runs with diurnally varying ozone show temperature differences of up to 3 K in the stratosphere, compared to those with climatological ozone (Sassi et al., 25). A prognostic ozone field also allows the modelling of feedbacks between radiation, chemistry and dynamics, and this is expected to improve forecasts, especially over longer timescales. However, no study has yet found a clear benefit in terms of forecast scores (e.g., Morcrette, 23). Finally, in order to simulate a good ozone distribution, models used in assimilation systems must be able to simulate stratospheric transport well; problems are often revealed when these models are confronted with real observations (e.g., Geer et al., 26b). Different approaches can be used for ozone data assimilation and the choice will vary depend upon the application. For studies of chemistry and transport, assimilation systems can be based on chemical transport models. These rely on operationally-produced analyses of wind and temperature (such as those from ECMWF) to advect chemical constituents. If chemistry is treated approximately, these models are extremely fast and can be used to assimilate many months of observations in a few days on a desktop computer. Including a detailed chemistry scheme, with dozens of constituents and hundreds of reactions, allows a more accurate simulation of the ozone distribution, but is slower. The CTM approach has been very popular for ozone analysis systems (e.g., Fisher and Lary, 1995; Khattatov et al., 2; Elbern and Schmidt, 21; Errera and Fonteyn, 21; Stajner et al., 21; Chipperfield et al., 22; Fierli et al., 22; Cathala et al., 23; Eskes et al., 23, 25a; El Amraoui et al., 24; Massart et al., 25; Segers et al., 25a; Wargan et al., 25). It is also possible to introduce a prognostic ozone field as a relatively straightforward upgrade to an existing NWP system (e.g., Struthers et al., 22; Dethof and Hólm, 24; Geer et al., 26b), though ozone assimilation then becomes part of a very large operational system, requiring a supercomputer. Much work is still required to confirm the proposed benefits (listed in the previous paragraph) of including ozone directly into NWP systems. One alternative approach is to couple CTMs, with a detailed description of chemistry, to GCM-based NWP systems, such that feedbacks between chemistry, dynamics and radiation can be maintained. Current operational satellite ozone observations include the Total Ozone Mapping Spectrometer (TOMS), measuring total column ozone, and the Solar Backscatter Ultraviolet (SBUV) instrument (e.g., Bhartia et al., 1996), which produces vertical profiles. ENVISAT, launched in 22, provides the instruments MIPAS (Michelson Interferometer for Passive Atmospheric Sounding), SCIAMACHY (Scanning Imaging Absorption Spectrometer for Atmospheric Chartography) and GOMOS (Global Ozone Monitoring by Occultation of Stars). Between them, these instruments measure many species beyond ozone, and vertical resolution is much improved over the operational instruments. For example, MI- PAS has roughly twice the vertical resolution of SBUV in the stratosphere (see e.g. Fig. 2, Wargan et al., 25). Dethof (23a) and Wargan et al. (25) show that ozone analyses are significantly better when MIPAS observations are assimilated, compared to experiments using only total column or low-vertical resolution observations like TOMS or SBUV. The ASSET project is based around assimilating the data from ENVISAT. The EOS-Aura satellite, launched in 24, has instruments with similar capabilities. Research instruments such as those on ENVISAT and Aura do, however, have a limited lifetime and data products are not always available quickly enough to be included in operational NWP schedules. Hence research satellite data is often best used Atmos. Chem. Phys., 6, , 26

3 A. J. Geer et al.: Intercomparison of ozone analyses 5447 Table 1. Principal features of the analysis systems and climatologies. See Sect. 2 for details. Name Type Winds Lon/Lat resolution (degrees) opera- ECMWF tional Heterogeneous ozone chemistry NWP.5/.5 (approx.) ECMWF MIPAS NWP.5/.5 (approx.) Levels from 1 to 1hPa Data assimilation scheme 21 4D-Var SBUV, GOME total columns, MIPAS from 7/1/ D-Var SBUV, GOME total columns, MIPAS throughout Cariolle v1.2 Cariolle v1.2 Ozone observations Ozone photochemistry T <195 K term T <195 K term DARC/Met Office NWP 3.75/ D-Var MIPAS Cariolle v1. Cold tracer KNMI TEMIS CTM ECMWF 3./2. 21 suboptimal KF KNMI SCIA- MACHY profiles CTM ECMWF 3./2. 21 suboptimal KF SCIAMACHY TOSOMI total columns SCIAMACHY profiles LINOZ Cariolle v1. Cold tracer Cold tracer BASCOE v3d24 CTM ECMWF 5./ D-Var MIPAS 57 species PSCBox BASCOE v3q33 CTM ECMWF 5./ D-Var MIPAS 57 species PSC parametrization MOCAGE- PALM/Cariolle MOCAGE- PALM/Reprobus CTM Arpege 2./2. 17 (to 5hPa) CTM Arpege 2./2. 17 (to 5hPa) 3-D- FGAT 3-D- FGAT MIMOSA CTM ECMWF 1./1. 14 suboptimal KF Juckes (26a) CTM ECMWF 2./2. (approx.) Logan/Fortuin/Kelder climatology 1 Direct inversion MIPAS Cariolle v2.1 T <195 K term MIPAS REPROBUS (Lefèvre et al., 1994) MIPAS None None MIPAS None None Carslaw et al. (1995) for re-analyses, and to help improve models and assimilation systems such that the operational observations may be assimilated more successfully. There have been a number of previous intercomparisons between the ozone distributions in chemistry-climate models (Austin et al., 23) and in CTMs (e.g., Bregman et al., 21; Roelofs et al., 23). These, and numerous other individual studies, have illustrated the importance of correctly modelling ozone transport. It is also well known that in CTMs driven by analysed wind fields, the stratospheric Brewer- Dobson circulation transports constituents far too quickly (Schoeberl et al., 23). This is the first time an intercomparison of ozone assimilation systems has been made. Historically, intercomparisons between GCMs have tried to standardise as many factors as possible (e.g. Gates, 1992; Pawson et al., 2). Compared to GCMs, there are many more areas where assimilation systems can diverge, such as in the assimilated observations (also their errors and the quality control applied to them), the assimilation technique, and the specification of background errors. This makes it much trickier to understand differences between experiments. Atmos. Chem. Phys., 6, , 26

4 5448 A. J. Geer et al.: Intercomparison of ozone analyses Since an intercomparison of assimilation systems is bound to be difficult, why not simply rely on model-only results? The reason is that assimilation systems often perform very differently to models. This can result in problems: well known examples include the spin-down of tropical moisture in the ECMWF system (Uppala et al., 25) and the excessive stratospheric Brewer-Dobson circulation produced by assimilated wind datasets. In contrast, some problems found in free-running models can be corrected, if not totally eliminated, by the regular insertion of observational data. For example, stratospheric tracer gradients become much more realistic in assimilation experiments compared to free model runs (e.g. Chipperfield et al., 22). Hence, data assimilation systems need to be separately validated. For this intercomparison, ozone analyses have been made for the period July to November 23, chosen because of the availability of good quality MIPAS data. This period included one of the largest ozone holes on record (e.g., Dethof, 23a), caused by relatively low temperatures in a fairly stable southern hemisphere (SH) polar vortex, which was destroyed by the usual top-down break-up during October and November (Lahoz et al., 26). Eleven analysis runs are included in the intercomparison, made using seven different systems, summarised in Table 1. As a reference, a climatology is also included in the intercomparison. Most systems assimilate MIPAS ozone observations, though KNMI analyses assimilate SCIAMACHY instead. Apart from the common time period, and (in most cases) the assimilation of MIPAS, we did not impose other standardisation. This was both for reasons of expediency and also to include as many analysis products as possible in the intercomparison. Both CTMs and GCMs are represented, and ozone chemistry may or may not be modelled. If included, it is done either by highly detailed reaction schemes or via a parametrization often known as a Cariolle scheme (e.g., Cariolle and Déqué, 1986; McLinden et al., 2; Mc- Cormack et al., 24). The Cariolle scheme is a linearisation of ozone photochemistry around an equilibrium state, using parameters derived from a more detailed model. Analyses are interpolated from their native resolution onto a common grid and then compared to independent ozone data from Halogen Occultation Experiment (HALOE), ozonesondes and TOMS, and to MIPAS. The common grid simplified the task of making the intercomparison, but it also introduced avoidable errors. Section 4.1 quantifies these errors; we do not believe they have much impact on our conclusions. Most of the analysis systems are focused on the stratosphere, but the scope of the comparison spans from the troposphere to the mesosphere. This is because many of the assimilation systems produce output over the whole vertical range. The quality of the analyses is expected to be poor in the troposphere but it is still useful to quantify and understand the errors in such regions. This paper introduces the intercomparison project, the method used, the independent data sources, and the analysis systems involved. It outlines many of the initial results and it draws initial conclusions on the various different methods used. There is not scope in this paper for detailed comparisons, such as between different types of chemistry schemes, or between 3-D-Var and 4-D-Var. These are anyway best performed in an experimental setting within a single assimilation system. However, the intercomparison provides a framework under which these results, and their significance, can be assessed by comparison to a variety of other assimilation approaches and systems. These more detailed results will be described in further papers. Most data, figures, and code are publicly available via the project website ( 2 Analyses Before describing in detail the analyses and climatologies in the intercomparison, we show examples of the ozone fields at 68 hpa on 31 August 23 (Fig. 1). Sunlight has started to return to high latitudes after the winter, triggering the depletion of ozone in a ring around the pole (see e.g., WMO, 23). Sunlight has not yet returned to the pole itself. The ring of higher ozone (3 to 5 ppmm) at about 45 S is the remainder of ozone that has descended throughout the SH high latitudes during the winter, from levels higher in the atmosphere where ozone amounts are greater. It is clear that at 68 hpa all the analyses show broadly similar and (from Sect. 5) realistic structures. Compared to the others, the KNMI SCIA- MACHY profile analyses have a bias; due to a lack of observations before October they are based principally on the free-running model. 2.1 ECMWF Ozone observations have been assimilated into the operational ECMWF analyses ( since April 22. During 23, GOME columns and SBUV profiles were assimilated, though in August and September 23, there was very limited availability of GOME data. MIPAS was assimilated operationally from from 7 October 23 until 25 March 24. Here we consider two datasets: (a) the operational analyses and (b) a dataset that includes assimilated MIPAS ozone throughout the July-November period, based on a pre-operational test suite before 7 October, and operational analyses after 7 October (Dethof, 23a). In all cases, the MIPAS data is version 4.59 of the Near Real Time product. Gross outliers in the MIPAS retrievals are rejected based on a comparison against the background ozone. Variational quality control is also applied (Andersson and Järvinen, 1999). The GCM in use when the analyses were made had a horizontal resolution of T511 ( 5 km) and 6 levels in the vertical, from the surface up to.1 hpa. Ozone was advected using a semi-lagrangian transport scheme. Ozone chemistry Atmos. Chem. Phys., 6, , 26

5 A. J. Geer et al.: Intercomparison of ozone analyses 5449 ECMWF operational ECMWF MIPAS DARC/Met Office UM KNMI TEMIS KNMI SCIA profiles BASCOE v3d24 BASCOE v3q33 MOCAGE-PALM Cariolle v Ozone /1-6kgkg MOCAGE-PALM Reprobus Juckes MIMOSA Logan/Fortuin/Kelder climatology Fig. 1. Ozone (ppmm) at 68 hpa in the southern hemisphere on 31 August 23, shown on a polar stereographic projection bounded by the equator. was parametrized with version 1.2 of the Cariolle scheme (Cariolle and De que, 1986; Dethof and Ho lm, 24), which includes a description of heterogeneous ozone depletion. Climatological ozone (Fortuin and Langematz, 1995), not prognostic ozone, was used for modelling heating rates. Data assimilation uses 4D-Var (e.g., Rabier et al., 2). Ozone is assimilated univariately, but it can still affect the dynamical analyses through the 4D-Var method (e.g., Riishøjgaard, 1996) and through the influence of ozone on the assimilation of temperature radiances. Background error correlations are calculated using an ensemble of analyses (Fisher, 23); background error variances are flow dependent. 2.2 DARC/Met Office The Met Office NWP system has recently been extended to allow the assimilation of ozone (Jackson and Saunders, 22; Jackson, 24) but ozone is not assimilated operationally. Here, MIPAS v4.61 ozone and temperature are assimilated in re-analysis mode, alongside all operational dynamical observations, using a stratosphere/troposphere version of the operational NWP system. The system is that described in Geer et al. (26b), but with a number of improvements to the GCM and no assimilation of HIRS (High resolution infrared radiation sounder) channel 9 ozone radiances. The assimilating GCM has a horizontal resolution of 3.75 longitude by 2.5 latitude and 5 levels in the vertical, from the surface to.1 hpa. It uses a new dynamical core (Davies et al., 25) which includes a semilagrangian transport scheme. This gives a better description of the Brewer-Dobson circulation than that seen in Geer et al. (26b). Ozone photochemistry is parametrized by v1. of the Cariolle and De que (1986) scheme. Improving on Geer et al. (26b), heterogeneous ozone chemistry is now parametrized, using a cold tracer scheme (Eskes et al., 23). Climatological ozone (Li and Shine, 1995), not the prognostic field, is used for modelling heating rates. Data assimilation uses 3-D-Var (Lorenc et al., 2). As for ECMWF, ozone is assimilated univariately, but 3-D-Var does not infer dynamical information, so the only effect of ozone on the dynamical analysis is through its influence on temperature radiance assimilation. Background error covariances are uniform for all latitudes and longitudes, and they are based on the ECMWF vertical covariances. As illustrated in Geer et al. (26b), the MIPAS ozone observations are subject to quality control, but with a lax threshold, so very few observations are rejected. Atmos. Chem. Phys., 6, , 26

6 545 A. J. Geer et al.: Intercomparison of ozone analyses 2.3 KNMI The Royal Netherlands Meteorological Institute (KNMI) operate a CTM (known as TM5) which has been used to assimilate SCIAMACHY ozone data. The CTM uses a subset of 44 of the ECMWF model levels, from the surface to.1 hpa, on a 3 longitude by 2 latitude grid. Data assimilation is done using a sub-optimal Kalman Filter (see e.g., Kalnay, 23), where the background error variances, but not the correlations, are advected as a tracer. Two different configurations are presented. The first configuration assimilates total column ozone from SCIAMACHY, retrieved at KNMI using the TOSOMI algorithm (Total Ozone retrieval scheme for SCIAMACHY based on the OMI DOAS algorithm, Eskes et al., 25b). The CTM is driven by ECMWF operational analyses of winds and temperatures. Ozone photochemistry is parametrized using the LINOZ scheme (McLinden et al., 2), a variant on Cariolle and Déqué (1986). Heterogeneous chemistry uses a cold tracer scheme. For assimilating total column observations, the vertical error correlations are set proportional to the vertical ozone profile. The system is very similar to that described in Eskes et al. (23). The second configuration assimilates ozone profiles (IFE v1.6) from the limb-sounding mode of SCIAMACHY. SCIA- MACHY limb profiles are mainly available for October and November 23; July to September is a free model run apart from a few assimilated profiles in August. The main uncertainty in the SCIAMACHY product is pointing, which has a vertical offset of 1 2 km (Segers et al., 25b). All profiles have been shifted in the vertical to get the best match with model forecasts prior to analysis. Ozone chemistry is parametrized using Cariolle and Déqué (1986) v1. and a cold-tracer scheme. The CTM is driven by ECMWF short range forecasts at 3 hourly intervals. The system is otherwise similar to that described in Segers et al. (25a). 2.4 BASCOE The Belgian Assimilation System of Chemical Observations from ENVISAT (BASCOE, is a 4D-Var assimilation system descended from that described in Errera and Fonteyn (21). Studies of the Antarctic and Arctic winter using the CTM of BASCOE can be found in Chabrillat et al. (26) 1 and Daerden et al. (26). MI- PAS v4.61 ozone (O 3 ), water vapour (H 2 O), nitric acid (HNO 3 ), nitric dioxide (NO 2 ), methane (CH 4 ) and nitrous oxide (N 2 O) are assimilated. Observations are subjected to an Optimal Interpolation Quality Check (OIQC, e.g. Gauthier et al., 23). In practice, the lowest MIPAS ozone ob- 1 Chabrillat, S. H., Van Roozendael, M., Daerden, F., Errera, Q., Hendrick, F., Bonjean, S., Wilms-Grabe, W., Wagner, T., Richter, A., and Fonteyn, D.: Quantitative assessment of 3-D PSCchemistry-transport models by simulation of GOME observations during the Antarctic winter of 22, in preparation, 26. servations in the ozone hole are rejected. Observations are also rejected if they fail a check for spurious vertical oscillations in the profile. The model includes 57 chemical species and 4 types of stratospheric PSC particles (ice; supercooled ternary solution, STS; nitric acid trihydrate, NAT; sulphuric acid tetrahydrate, SAT) with a full description of stratospheric chemistry and microphysics of PSCs. All chemical species are advected and interact through 143 gas-phase reactions, 48 photolysis reactions and 9 heterogeneous reactions. To allow for calculating transport of PSCs, size distributions of each type are discretized using 36 bins from.2 to 36 µm. PSC microphysics is described by the PSCBox scheme (Larsen et al., 2). In order to improve agreement with MIPAS ozone, O 2 photolysis rates were multipled by This version is referred to as v3d24. Based on early results of this intercomparison, a new version of BASCOE, v3q33, was produced. Among the changes, v3q33 replaces the full PSC calculation by a parametrization that defines (1) surface area density of ice and NAT when their occurrence is possible and (2) the loss of HNO 3 and H 2 O due to sedimentation (Chabrillat et al., 26 1 ). Ice PSCs are supposed to exist in the winter/spring polar regions at any grid point where the temperature is colder than 186 K, and NAT PSCs at any grid point where the temperature is colder than 194 K. The surface area density is set to 1 6 cm 2 /cm 3 in the first case and 1 7 cm 2 /cm 3 in the second. Additionally in v3q33, O 2 photolysis rates are no longer scaled; this reduces the bias against HALOE but increases it against MIPAS (see Sect. 5.1). Finally, the Arakawa A grid of v3d24 was replaced by a C grid (see e.g., Kalnay, 23) in v3q33. The CTM is driven by ECMWF operational analyses of winds and temperatures, and uses a subset of 37 of the ECMWF model levels, from the surface to.1 hpa, on a 5 longitude by 3.75 latitude grid. Data assimilation is done using 4D-Var. The background error standard deviation is set as 2 of the background ozone amount. Though there are no off-diagonal elements in the background error covariances (i.e. no vertical or horizontal correlations), information from MIPAS observations is still spread through the observation operator, as in other systems. Here, it averages the 8 grid points surrounding the measurement point, and the relatively broad horizontal resolution of the grid also helps to spread the information. 2.5 Météo-France/CERFACS The Météo-France/CERFACS assimilation system is based upon the 3-D CTM MOCAGE and the PALM software (Massart et al., 25). MIPAS v4.61 ozone data are assimilated, but not poleward of 8 of latitude. The PALM framework is particularly versatile, as both the CTM degree of sophistication (for instance, the number of chemical tracers involved, the physical or chemical Atmos. Chem. Phys., 6, , 26

7 A. J. Geer et al.: Intercomparison of ozone analyses 5451 parametrizations, the horizontal and vertical geometries) and the data assimilation technique can be changed easily; this makes the MOCAGE-PALM system a useful platform for sensitivity studies in chemical data assimilation. MOCAGE is a flexible tropospheric and stratospheric 3- D CTM developed at Météo-France, offering several configurations of varying computational costs. Two separate configurations are examined here. The first uses linear ozone chemistry, with v2.1 of the Cariolle and Déqué (1986) scheme. The second includes a detailed representation of stratospheric and upper tropospheric chemistry, based upon the REPROBUS chemical scheme (Lefèvre et al., 1994), which comprises 55 transients and species and takes into account heterogeneous chemistry on polar stratospheric clouds (Carslaw et al., 1995; Lefèvre et al., 1998). The REPROBUS chemistry version of MOCAGE has already been used for UTLS assimilation studies (Cathala et al., 23). A more comprehensive version of MOCAGE, with comprehensive tropospheric chemistry, is run daily in operational mode at Météo-France for chemical weather and air quality applications (Dufour et al., 24, see daily global forecasts at MOCAGE relies on a semi-lagrangian advection scheme (Josse et al., 24). For the experiments presented here, MOCAGE has a 2 by 2 horizontal resolution and 47 hybrid sigma/pressure levels extending from the surface up to 5 hpa. The meteorological forcings are Météo-France ARPEGE operational meteorological analyses of pressure, winds, temperature and humidity (Courtier et al., 1991), available every 6 h. Any assimilation algorithm can be seen as a sequence of elementary operations or elementary components that can exchange data (Lagarde et al., 21). Based on this idea, the CERFACS PALM software ( palm) manages the dynamic launching of the coupled components (forecast model, algebra operators, I/O of observational data) and the parallel data exchanges. The MOCAGE- PALM assimilation system is set up here in a 3-D-FGAT configuration (3-D First Guess at Assimilation Time, Fisher and Andersson, 21). As a first approximation, background error standard deviations are prescribed as 2 of the background ozone amount. In order to spread assimilation increments spatially, horizontal background error correlations are modelled using a generalized diffusion operator (Weaver and Courtier, 21), with a length-scale of 4 ; no vertical background error correlations are considered. 2.6 MIMOSA MIMOSA (Modèle Isentrope de transport Mésoéchelle de l Ozone Stratosphérique par Advection) is a CTM driven by ECMWF operational winds and temperatures (Fierli et al., 22). MIPAS v4.61 ozone data are assimilated. There is no quality control; all observations are included. There are 16 isentropic levels from 335 K to 165 K, approximately spanning the stratosphere ( 2 hpa to 1 hpa) and a 1 by 1 latitude-longitude grid. Advection is semi-lagrangian. The model includes neither ozone chemistry nor cross-isentropic transport. Data assimilation is done using a sub-optimal Kalman Filter with advected background error variances, and uses the Physical Space Assimilation System method (PSAS, e.g. Kalnay, 23). Background error correlations are flow dependent and anisotropic, specified in terms of distance and the potential vorticity (PV) field. The model error covariance (Q) is diagonal, and proportional to the ozone amount, x, e.g. Q=(qx) 2 where q=.24 day 1 and has been tuned using χ 2 tests. ECMWF operational temperature and pressure fields are used to interpolate these isentropic analyses onto pressure levels for this study. 2.7 Juckes These are analyses produced by a direct inversion method (Juckes, 26a,b), which assimilates months of MIPAS v4.61 ozone data by making a single iterative solution. The physical constraint is based on an isentropic transport equation. Rather than discretising the predictive equations (which would give a CTM), the product of these equations with their adjoint is discretised. The resulting self-adjoint system of equations is solved with a multigrid relaxation algorithm. This is equivalent to solving the Kalman Smoother (e.g. Rodgers, 2) with fully advected background error covariances. However, this technique avoids the need to represent the background error covariance matrix explicitly. Ozone transport is driven by ECMWF operational winds and temperatures, on 13 isentropic levels from 38 K to 3 K. In the horizontal, a binary thinned latitude-longitude spherical grid is used, giving approximately 2 by 2 resolution. The model error covariance (Q) is diagonal, with a constant value of.2 ppmv 2 /day 2. The model includes neither ozone chemistry nor cross-isentropic transport. As for MIMOSA, ECMWF pressure and temperature fields are used for interpolation onto pressure levels in this study. 2.8 Climatology To contrast with the assimilated ozone fields, we include a climatology-derived product in the comparison. As a minimum, we would expect the analyses to do better than climatology. We combine the Logan (1999) tropospheric ozone climatology with the Fortuin and Kelder (1998) stratospheric ozone climatology. In each case, the climatologies are resolved on a monthly basis. The Logan (1999) climatology uses ozonesonde, surface in-situ data and the TOMS/Stratospheric Aerosol and Gas Experiment (SAGE) tropospheric residual, to produce a partly 3-D and partly 2-D climatology on 13 levels from 1 hpa to 1 hpa, covering latitudes from 89 S to 89 N. Atmos. Chem. Phys., 6, , 26

8 5452 A. J. Geer et al.: Intercomparison of ozone analyses Pressure /hpa (a) (b) ECMWF operational ECMWF MIPAS DARC/Met Office UM KNMI SCIA profiles KNMI TEMIS BASCOE v3d24 BASCOE v3q33 MOCAGE-PALM Cariolle v2.1 MOCAGE-PALM Reprobus Juckes MIMOSA Logan/Fortuin/Kelder climatology B error std. dev. / B error std. dev. / Fig. 3. Key to the analyses. Typically only a subset of these are shown in any one figure. Fig. 2. Standard deviation of ozone background error, given as a percentage relative to climatological ozone, averaged for the period 7 October 23 to 31 October 23 in the regions (a) 3 S to 3 N and (b) 6 S to 9 S, for the analysis systems of ECMWF, DARC, BASCOE and MOCAGE-PALM (see key in Fig. 3). MIPAS error standard deviations, averaged for the same region and time period, are shown by the black line. The Fortuin and Kelder (1998) climatology uses ozonesondes, SBUV and TOMS total ozone, from 198 to 1991, to produce a 2-D (latitude-pressure) climatology with 19 levels from 1 hpa to.3 hpa and covering latitudes from 8 S to 8 N. An ozone field was created on the intercomparison common grid, daily at Z and 12Z, by interpolating the climatologies linearly in time, and treating the climatologies as representative of the 15th of each month. Beyond the northern and southern limits of the climatologies, horizontal extrapolation was done at constant value. Logan (1999) values were taken for levels at 15 hpa and below, and Fortuin and Kelder (1998) above, up to.3 hpa. Figure 1 shows that this results in a zonal distribution which, as expected, does not represent the synoptic features in the ozone field. 2.9 Comparison of ozone background errors The background error covariance matrix (e.g., Kalnay, 23) is important in determining the weight given to observations in data assimilation. In general, at the observation point, more weight is given to the model as the background error standard deviation becomes smaller compared to the observation error standard deviation. However, the spreading of information away from the observation point is determined by the background error correlations, any observation error correlations (not usually considered), and the observation operator. Here, only the DARC and ECMWF systems include vertical correlations in the background errors. The general impact of observations on the system will also depend on how many observations are rejected by quality control. We examine the background error standard deviation (i.e. the square root of the diagonal of the background error covariance matrix) from a number of the analysis systems, and compare it to the MIPAS ozone observation error standard deviation (Fig. 2). In each case these have been normalised by the climatological ozone amount (see Sect. 4 for the method). The comparison is done to illustrate the varied approaches to background error modelling, and to give some indication of the weight assigned to observations in the different systems. As already noted, however, many other factors affect the observations weight in the final analysis. Figure 2 shows large differences in the ozone background error standard deviations assumed in the assimilation systems. In the mid and upper stratosphere (levels above 3 hpa), DARC and ECMWF background error standard deviations are less than 5 of the ozone field. These are ultimately derived from an ensemble of analyses (Fisher, 23); the small standard deviations in the upper stratosphere and mesosphere are most likely due to the strong control of the ozone field by parametrized ozone photochemistry schemes at these levels. The BASCOE and (to 5 hpa) Météo- France/CERFACS systems have background error standard deviations of typically 2, based on ad-hoc methods. At these levels the DARC and ECMWF analyses are likely to give less weight to observations. MIPAS observation errors are markedly larger at the tropical tropopause, and all the systems are likely to give relatively lower weight to MIPAS observations here than in the rest of the stratosphere. The specification of background errors is known to be one of the most difficult parts of any assimilation system. From Fig. 2 and the details given earlier, we see that ozone assimilation systems use very varied and often ad-hoc approaches to modelling background error covariances. Despite this, we later find that the quality of the analyses is often similar: any Atmos. Chem. Phys., 6, , 26

9 A. J. Geer et al.: Intercomparison of ozone analyses 5453 differences coming from the background errors are too small to be identified in an intercomparison. This is certainly an area where more studies are needed. 1 8 MIPAS coverage Jul 3 Ozone observations 3.1 MIPAS MIPAS observations are used in two ways in this paper. First, MIPAS ozone observations have been assimilated into all the analyses, except those from KNMI, which assimilated SCIA- MACHY instead. The exact set of MIPAS observations included in each system will vary, depending upon the quality control procedures used. Second, in order to calculate statistics for the intercomparison, we have compared a fixed set of MIPAS observations to all the analyses including those from KNMI. MIPAS is an interferometer for measuring infrared emissions from the atmospheric limb (Fischer and Oelhaf, 1996). MIPAS operational data are available between July 22 and March 24, after which instrument problems meant it could only be used on an occasional basis. The operational measurements were made along 17 discrete lines-of-sight in the reverse of the flight direction of ENVISAT, with tangent heights between 8 km and 68 km. The vertical resolution was 3 km and the horizontal resolution was 3 km along the line of sight. ENVISAT follows a sun-synchronous polar orbit, allowing MIPAS to sample globally, and to produce up to 1 atmospheric profiles per day. Coverage is quite uniform in time and there are only minor variations with latitude (see Fig. 4). From the infrared spectra, ESA retrieved profiles of pressure, temperature, ozone, water vapour, HNO 3, NO 2, CH 4 and N 2 O at up to 17 tangent points (ESA, 24; Raspollini et al., 26). MIPAS version 4.61 data, reprocessed offline, is used throughout this work, except in the ECMWF assimilation runs, where the Near Real Time v4.59 product was used. The differences between v4.59 and v4.61 processors are minor. MIPAS ozone appears unbiased when compared to independent data except in the lower stratosphere where a small positive bias has been noted (Dethof, 23a,b, 24; Fischer and Oelhaf, 24; Wargan et al., 25; Geer et al., 26b). However, a comparison against ozonesondes using the MI- PAS averaging kernels, but only a limited number of colocations, identified no bias (Migliorini et al., 24). In Sect. 5.3 we find a positive bias of order 5 in the upper stratosphere with respect to HALOE, increasing to roughly 1 with respect to sonde and HALOE in the lower stratosphere. The official MIPAS validation papers are currently in preparation. Typically, assimilation systems produce observation minus first guess (O-F) statistics that are used for monitoring biases between the observations and the models, and checking that statistical assumptions are valid in the assimilation Count Aug Sep Oct Nov Latitude Fig. 4. Number of MIPAS profiles used for validation, by latitude (in 1 bins) and by month, for July to November 23. Histograms for different months have been staggered by an interval of 2 counts. algorithm (e.g., Talagrand, 23; Stajner et al., 24). We did not use the O-F statistics produced by the assimilation systems themselves but instead calculated statistics from a comparison between the common-gridded analyses and a fixed set of MIPAS observations. This was done to avoid reading a plethora of different file formats, to ensure that a fixed set of observations was used, and to be able to calculate statistics for those analyses which did not include MIPAS observations. Section 4.1 quantifies the errors resulting from the use of the common grid. The fixed set of MIPAS observations included all those supplied in the ESA data files except those that failed screening criteria developed during data assimilation experiments at DARC (Lahoz et al., 26). To calculate statistics of the difference between analysis and MIPAS, the MIPAS retrievals are treated as point measurements and the analyses are interpolated to the MIPAS retrieval points linearly in ln(p ) in the vertical. The paired differences are then binned to the nearest pressure level on the intercomparison grid. It is well known that comparison in terms of radiances, or the use of averaging kernels (Rodgers, 2; Migliorini et al., 24), produces a better representation of the information content of the retrievals; these methods are increasingly used in calibration and validation activities. But here, no assimilation system uses MIPAS radiances or an averaging kernel representation, so it is the bias in MI- PAS retrievals, treated as a point values, that is important. 3.2 Ozonesondes Ozonesondes are used as independent data to validate the analyses. Profiles have been obtained from the World Ozone and Ultraviolet Radiation Data Centre (WOUDC, Atmos. Chem. Phys., 6, , 26

10 5454 A. J. Geer et al.: Intercomparison of ozone analyses 25 Sonde coverage Jul 1 2 Aug Count 15 1 Sep Oct 1 5 Nov Latitude Pressure /hpa 1 Fig. 5. Number of ozonesonde profiles used for validation, by latitude (in 1 bins) and by month, for July to November 23. Histograms for different months have been staggered by an interval of 5 counts. Southern Hemisphere Additional Ozonesondes project (SHADOZ, shadoz/, Thompson et al., 23a,b) and the Network for the Detection of Stratospheric Change (NDSC, ndsc.ncep.noaa.gov/). We use ozonesonde ascents from 42 locations, not including the Indian stations, and comprising mostly Electrochemical Concentration Cell (ECC) types, with five locations using Carbon-Iodide sondes and one location using Brewer-Mast sondes. We approach this dataset in the knowledge that it may be somewhat heterogeneous, both in the sonde types used, but also in the correction factors applied to the data, and in the operating procedures at each site. See Komhyr et al. (1995) and Thompson et al. (23a) for more discussion of the importance of these techniques and procedures. However, we believe this heterogeneity is worth accepting in order to gain the widest global coverage. The number of sonde ascents available to this intercomparison, and their latitudinal and temporal coverage, are summarised in Fig. 5. Sondes typically make measurements from the surface to around the 1 hpa level. Total error for ECC sondes is estimated to be within 7 to +17 in the upper troposphere, ±5 in the lower stratosphere up to 1 hpa and 14 to +6 at 4 hpa (Komhyr et al., 1995). Errors are higher in the presence of steep ozone gradients and where ozone amounts are low. In order to compare sonde profiles, with a relatively high vertical resolution, to the analyses on the intercomparison common grid, the sonde profiles are averaged over a layer bounded by the half-way points (calculated linearly) between the common pressure levels. For example, analyses at 1 hpa are compared to the mean of any ozonesonde pro ozone /mpa Fig. 6. Example of comparison between sonde ozone at full resolution (black line), layer-averaged (black triangles) and the analyses (key in Fig. 3) at 11:36 UTC on 24 September 23 at Legionowo (21. E, 52.4 N). file points between 125 hpa and 84 hpa. Especially within the polar vortex, sondes may drift long distances during their ascent, but tracking information is not always supplied. We disregard the horizontal movement of sondes and assign the measurement position as the launch longitude and latitude. Figure 6 gives an example of the intercomparison, showing both the full-resolution sonde profile and the layeraverages used to calculate statistics, alongside a number of different analyses. We will return to this figure in Sect HALOE HALOE (Russell et al., 1993) uses solar occultation to derive atmospheric constituent profiles. HALOE is used here as independent data for validation. Figure 7 shows the coverage available. The nature of the solar occultation technique makes the data sparse in time and space, with about 15 observations per day at each of two latitudes. The horizontal resolution is 495 km along the orbital track and the vertical resolution is about 2.5 km. Atmos. Chem. Phys., 6, , 26

11 A. J. Geer et al.: Intercomparison of ozone analyses 5455 We use an updated version 19 product, screened for cloud using the algorithm of Hervig and McHugh (1999), and available from the HALOE website ( nasa.gov/). We found that, compared to the previously available version 19, the one with cloud screening substantially improved the quality of results in this intercomparison around the tropical tropopause. Aside from the cloud screening, version 19 ozone retrievals are nearly identical to those of v18, and above the 12 hpa level they agree with ozonesonde data to within 1 (Bhatt et al., 1999). Below this level, profiles can be seriously affected by the presence of aerosols and cirrus clouds. HALOE profiles are supplied on 271 levels with very close vertical spacing, but vertical variation is smooth due to the much broader vertical resolution of the instrument. Hence, to compare to the analyses on common pressure levels, HALOE is simply interpolated between the nearest two of the 271 levels. Longitudes and latitudes vary with height in HALOE profiles but for this comparison, those at 1 hpa are taken to be representative of all levels. 3.4 TOMS The Total Ozone Mapping Spectrometer (TOMS) measures backscattered ultraviolet radiances with high horizontal resolution (38 km by 38 km) and daily near-global coverage. There are small gaps between orbital coverage bands near the equator. During the intercomparison period, due to the lack of sunlight at very high latitudes, there is no data in July, August and September in the southern high latitudes; the same for October and November in the north. TOMS is not assimilated in any of the analyses evaluated here. We use version 8 of the level 3 total column ozone product, which is a daily composite of binned observations. Version 8 has partial corrections for calibration problems in the post-2 TOMS data from the Earth Probe satellite, and improved retrievals under extreme conditions (high observation angles, in the Antarctic, aerosol loading) compared to v7 (McPeters, personal communication, 24). A full validation of TOMS v8 has not yet been published, but v7 uncertainties were estimated as about 2 for the random errors, 3 for the absolute errors and somewhat more at high latitudes due to the higher zenith angle (McPeters et al., 1998). 4 Method All analyses were interpolated from their native resolutions onto a common grid. This was done to make the intercomparison task easier, but with the knowledge that it would introduce a small extra source of error, quantified in Sect The resolution of the common grid was determined by the need to minimise storage requirements whilst not losing important geophysical variability in time or space, and so minimising colocation error when comparing with independent Count HALOE coverage Jul Aug Sep Oct Nov Latitude Fig. 7. Number of HALOE profiles used for validation, by latitude (in 1 bins) and by month, for July to November 23. Histograms for different months have been staggered by an interval of 3 counts. data. Based on the results of sensitivity tests (Sect. 4.1), the choice of a 3.75 longitude by 2.5 latitude grid, 37 fixed pressure levels, and twice daily analyses (Z and 12Z) appears to be a reasonable compromise. Pressure levels are 6 per decade between.1 hpa and 1 hpa (as used on the Upper Atmosphere Research Satellite (UARS) project). Below this, there are levels at 15, 2 hpa, and so on every 5 hpa down to 1 hpa. All comparisons against independent data, except those against TOMS (see Sect. 4.1), were made using analyses on the common grid. All vertical interpolations were done linearly in ln(p ) (where P is pressure) and all horizontal interpolation, bilinearly in longitude and latitude. Statistics were built up from the difference between analyses and observations. In this paper, statistics were binned in the regions referred to here as the southern and northern high latitudes (9 S to 6 S and 6 N to 9 N, respectively), the southern and northern midlatitudes (6 S to 3 S and 3 N to 6 N respectively) and the tropics (3 S to 3 N). Statistics were binned monthly; also for the entire period 18 August 23 to 3 November 23 (before 18 August 23, the DARC analyses were not adequately spun up). Ozone amounts vary by many orders of magnitude through the atmosphere. Units of partial pressure emphasise the UTLS; units of mixing ratio emphasise the mid and upper stratosphere. In order to give approximately equal weight through the atmosphere, statistics were normalised with respect to climatology, and displayed as a percentage. As an example, for a particular bin (e.g. July in the tropics at 1 hpa), where i runs over all n observations in this bin, the percentage mean difference d, between analysis interpolated Atmos. Chem. Phys., 6, , 26